[Bug target/116413] New: [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[Bug target/116413] New: [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[schwab@linux-m68k.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

            Bug ID: 116413
           Summary: [LRA] [M68K] ICE: unrecognized insn in
                    lra_set_insn_recog_data, at lra.cc:1036
           Product: gcc
           Version: 15.0
            Status: UNCONFIRMED
          Keywords: ice-on-valid-code, ra
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: schwab@linux-m68k.org
            Blocks: 113939
  Target Milestone: ---

With LRA enabled:

$ gcc/xgcc -B gcc/ -mcpu=5235 -Os
../gcc/gcc/testsuite/gcc.target/m68k/pr64461.c -S
during RTL pass: reload
../gcc/gcc/testsuite/gcc.target/m68k/pr64461.c: In function
‘rtems_rfs_block_map_indirect_alloc’:
../gcc/gcc/testsuite/gcc.target/m68k/pr64461.c:16:1: internal compiler error:
in lra_set_insn_recog_data, at lra.cc:1036
   16 | }
      | ^
0x1bf2c7e internal_error(char const*, ...)
        ../../gcc/gcc/diagnostic-global-context.cc:491
0x6734a1 fancy_abort(char const*, int, char const*)
        ../../gcc/gcc/diagnostic.cc:1772
0xbfc0df lra_set_insn_recog_data(rtx_insn*)
        ../../gcc/gcc/lra.cc:1034
0xbfca56 lra_get_insn_recog_data(rtx_insn*)
        ../../gcc/gcc/lra-int.h:503
0xbfca56 setup_sp_offset
        ../../gcc/gcc/lra.cc:1877
0xbfe415 lra_process_new_insns(rtx_insn*, rtx_insn*, rtx_insn*, char const*)
        ../../gcc/gcc/lra.cc:1925
0xc15ca6 curr_insn_transform
        ../../gcc/gcc/lra-constraints.cc:4868
0xc17286 lra_constraints(bool)
        ../../gcc/gcc/lra-constraints.cc:5496
0xbff57a lra(_IO_FILE*, int)
        ../../gcc/gcc/lra.cc:2442
0xbad357 do_reload
        ../../gcc/gcc/ira.cc:5976
0xbad357 execute
        ../../gcc/gcc/ira.cc:6164


Referenced Bugs:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113939
[Bug 113939] Switch m68k to LRA

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[schwab@linux-m68k.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #1 from Andreas Schwab <schwab@linux-m68k.org> ---
(insn 20 17 21 2 (set (reg:SI 43)
        (plus:SI (ashift:SI (reg/v:SI 40 [ b ])
                (const_int 2 [0x2]))
            (reg/v/f:SI 39 [ buffer ])))
"../../gcc/gcc/testsuite/gcc.target/m68k/pr64461.c":15:25 -1
     (nil))

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[schwab@linux-m68k.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #2 from Andreas Schwab <schwab@linux-m68k.org> ---
This is generated without LRA:

(insn 22 17 9 2 (set (reg:SI 9 %a1)
        (plus:SI (plus:SI (mult:SI (reg/v:SI 0 %d0 [orig:40 b ] [40])
                    (const_int 4 [0x4]))
                (reg/v/f:SI 9 %a1 [orig:38 map ] [38]))
            (const_int 3 [0x3])))
"../../gcc/gcc/testsuite/gcc.target/m68k/pr64461.c":15:25 407 {*lea}
     (nil))
(insn 9 22 12 2 (set (mem:QI (plus:SI (plus:SI (mult:SI (reg/v:SI 0 %d0
[orig:40 b ] [40])
                        (const_int 4 [0x4]))
                    (reg/v/f:SI 8 %a0 [orig:39 buffer ] [39]))
                (const_int 3 [0x3])) [0 *_5+0 S1 A8])
        (mem:QI (reg:SI 9 %a1) [1 map_8(D)->blocks[b_9(D)]+3 S1 A8]))
"../../gcc/gcc/testsuite/gcc.target/m68k/pr64461.c":15:25 63 {*m68k.md:1137}
     (nil))

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[schwab@linux-m68k.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #3 from Andreas Schwab <schwab@linux-m68k.org> ---
The issue is that LRA canonicalizes MULT to ASHIFT (outside of MEM), but
m68k_decompose_index only recognizes MULT, so that *lea fails to match.

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[schwab@linux-m68k.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #4 from Andreas Schwab <schwab@linux-m68k.org> ---
Created attachment 58956
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=58956&action=edit
Accept ASHIFT like MULT

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[schwab@linux-m68k.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #5 from Andreas Schwab <schwab@linux-m68k.org> ---
With that patch, *lea is used in more places and LRA fails to remove the inner
SUBREG.

../../../gcc/libgcc/libgcov-profiler.c: In function ‘__gcov_interval_profiler’:
../../../gcc/libgcc/libgcov-profiler.c:46:1: error: insn does not satisfy its
constraints:
   46 | }
      | ^
(insn 41 65 66 6 (set (reg/f:SI 8 %a0 [orig:44 _14 ] [44])
        (plus:SI (ashift:SI (subreg:SI (reg/v:DI 0 %d0 [orig:47 delta ] [47])
4)
                (const_int 3 [0x3]))
            (reg:SI 9 %a1))) "../../../gcc/libgcc/libgcov-profiler.c":45:13 407
{*lea}
     (nil))
during RTL pass: postreload

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[schwab@linux-m68k.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #6 from Andreas Schwab <schwab@linux-m68k.org> ---
Created attachment 58957
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=58957&action=edit
_gcov_interval_profiler.i

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[schwab@linux-m68k.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #7 from Andreas Schwab <schwab@linux-m68k.org> ---
Comment #5 was built _without_ LRA.

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[matz at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

Michael Matz <matz at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |matz at gcc dot gnu.org

--- Comment #8 from Michael Matz <matz at gcc dot gnu.org> ---
Yeah, reload doesn't expect ASHIFTs within operands here.  The below would fix
that:

diff --git a/gcc/final.cc b/gcc/final.cc
index eb9e065d9f0..5d911586de5 100644
--- a/gcc/final.cc
+++ b/gcc/final.cc
@@ -3146,6 +3146,7 @@ walk_alter_subreg (rtx *xp, bool *changed)
     case PLUS:
     case MULT:
     case AND:
+    case ASHIFT:
       XEXP (x, 0) = walk_alter_subreg (&XEXP (x, 0), changed);
       XEXP (x, 1) = walk_alter_subreg (&XEXP (x, 1), changed);
       break;

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[rsandifo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

Richard Sandiford <rsandifo at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rsandifo at gcc dot gnu.org

--- Comment #9 from Richard Sandiford <rsandifo at gcc dot gnu.org> ---
(In reply to Michael Matz from comment #8)
> Yeah, reload doesn't expect ASHIFTs within operands here.  The below would
> fix
> that:
> 
> diff --git a/gcc/final.cc b/gcc/final.cc
> index eb9e065d9f0..5d911586de5 100644
> --- a/gcc/final.cc
> +++ b/gcc/final.cc
> @@ -3146,6 +3146,7 @@ walk_alter_subreg (rtx *xp, bool *changed)
>      case PLUS:
>      case MULT:
>      case AND:
> +    case ASHIFT:
>        XEXP (x, 0) = walk_alter_subreg (&XEXP (x, 0), changed);
>        XEXP (x, 1) = walk_alter_subreg (&XEXP (x, 1), changed);
>        break;
LGTM FWIW.

The different canonicalisation rules around ASHIFT/MULT in memories vs outside
memories is an unfortunate wart, but what LRA is doing for lea is the correct
behaviour.

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[schwab@linux-m68k.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #10 from Andreas Schwab <schwab@linux-m68k.org> ---
$ ./xgcc -B . ../../gcc/gcc/testsuite/gcc.c-torture/compile/opout.c -O
-std=gnu89 -mlra -S
during RTL pass: reload
../../gcc/gcc/testsuite/gcc.c-torture/compile/opout.c: In function ‘z’:
../../gcc/gcc/testsuite/gcc.c-torture/compile/opout.c:12:1: internal compiler
error: in decompose_normal_address, at rtlanal.cc:6699
   12 | }
      | ^

(plus:SI (plus:SI (ashift:SI (plus:SI (reg/f:SI 24 %argptr)
                (const_int -4 [0xfffffffffffffffc]))
            (const_int 1 [0x1]))
        (reg/f:SI 41 [ _6 ]))
    (const_int 20 [0x14]))

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[rsandifo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #11 from Richard Sandiford <rsandifo at gcc dot gnu.org> ---
I can't reproduce this with m68k-elf.  Do you have any local changes beyond
making TARGET_LRA_P return true?

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[schwab@linux-m68k.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #12 from Andreas Schwab <schwab@linux-m68k.org> ---
Which cpu are you targeting?  m68k-linux defaults to 68020.

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[schwab@linux-m68k.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #13 from Andreas Schwab <schwab@linux-m68k.org> ---
And of course this is with #c4 applied.

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[rsandifo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #14 from Richard Sandiford <rsandifo at gcc dot gnu.org> ---
Created attachment 58967
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=58967&action=edit
Patch for the decompse_mem_address ICE

Thanks.  Can you try the attached patch?

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[schwab@linux-m68k.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #15 from Andreas Schwab <schwab@linux-m68k.org> ---
That fixes the ICE from #10 and no further regression in gcc.c-torture.

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[schwab@linux-m68k.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #16 from Andreas Schwab <schwab@linux-m68k.org> ---
$ gcc/xgcc -B gcc/ -fpic ../gcc/gcc/testsuite/gcc.dg/pr52139.c -S -mlra -O 
-fno-tree-dominator-opts -fno-tree-fre
../gcc/gcc/testsuite/gcc.dg/pr52139.c: In function ‘foo’:
../gcc/gcc/testsuite/gcc.dg/pr52139.c:50:1: error: unrecognizable insn:
   50 | }
      | ^
(insn 9 49 50 4 (set (reg:HI 33)
        (mem/u/c:HI (plus:SI (ashift:SI (reg/v:SI 32 [ a ])
                    (const_int 1 [0x1]))
                (label_ref:SI 11)) [0  S2 A8]))
"../gcc/gcc/testsuite/gcc.dg/pr52139.c":12:3 -1
     (insn_list:REG_LABEL_OPERAND 11 (nil)))
during RTL pass: vregs

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[rsandifo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #17 from Richard Sandiford <rsandifo at gcc dot gnu.org> ---
Hmm, but if the ICE is coming from vregs then it doesn't sound like it's
related to LRA.  vregs is the first RTL pass to run.

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[matz at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #18 from Michael Matz <matz at gcc dot gnu.org> ---
This last problem is a pre-existing one in the backend.  It accepts bare
label_ref plus operands, but it does so also in flag_pic mode.  There it
isn't correct (because it implies another memory reference).  This problem
was masked earlier because expand generates the address reference to the
jump table with a ASHIFT, which wasn't accepted formerly by the backend.
That non-acceptance lead to expand decomposing everything in the address
to registers, including the bare label_ref, which then ultimately went the
correct ways.

The LRA patches now _do_ accept the ASHIFT parts (and that's okay), but this
leads to an overall acceptance of '(plus (ashift (reg) (1)) (label_ref))'
and that's not acceptable (ahem).  Solution: don't accept a bare label_ref
as plus operand if flag_pic.  See below:

@@ -2247,6 +2275,7 @@ m68k_decompose_address (machine_mode mode, rtx x,
      references.  It seems unlikely that we would ever generate indexed
      accesses to unplaced labels in other cases.  */
   if (GET_CODE (x) == PLUS
+      && !flag_pic
       && m68k_jump_table_ref_p (XEXP (x, 1))
       && m68k_decompose_index (XEXP (x, 0), strict_p, address))
     {

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #19 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The trunk branch has been updated by Richard Sandiford <rsandifo@gcc.gnu.org>:

https://gcc.gnu.org/g:708ee71808ea61758e73d0e36274b4194b28576a

commit r15-3213-g708ee71808ea61758e73d0e36274b4194b28576a
Author: Richard Sandiford <richard.sandiford@arm.com>
Date:   Tue Aug 27 09:48:28 2024 +0100

    Handle arithmetic on eliminated address indices [PR116413]

    This patch fixes gcc.c-torture/compile/opout.c for m68k with LRA
    enabled.  The test has:

    ...
    z (a, b)
    {
      return (int) &a + (int) &b + (int) x + (int) z;
    }

    so it adds the address of two incoming arguments.  This ends up
    being treated as an LEA in which the "index" is the incoming
    argument pointer, which the LEA multiplies by 2.  The incoming
    argument pointer is then eliminated, leading to:

    (plus:SI (plus:SI (ashift:SI (plus:SI (reg/f:SI 24 %argptr)
                    (const_int -4 [0xfffffffffffffffc]))
                (const_int 1 [0x1]))
            (reg/f:SI 41 [ _6 ]))
        (const_int 20 [0x14]))

    In the address_info scheme, the innermost plus has to be treated
    as the index "term", since that's the thing that's subject to
    index_reg_class.

    gcc/
            PR middle-end/116413
            * rtl.h (address_info): Update commentary.
            * rtlanal.cc (valid_base_or_index_term_p): New function, split
            out from...
            (get_base_term, get_index_term): ...here.  Handle elimination
PLUSes.

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #20 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Michael Matz <matz@gcc.gnu.org>:

https://gcc.gnu.org/g:359209bdc7245f8768b5044acded8509545e4990

commit r15-3219-g359209bdc7245f8768b5044acded8509545e4990
Author: Michael Matz <matz@suse.de>
Date:   Thu Aug 22 17:03:56 2024 +0200

    final: go down ASHIFT in walk_alter_subreg

    when experimenting with m68k plus LRA one of the
    changes in the backend is to accept ASHIFTs (not only
    MULT) as scale code for address indices.  When then not
    turning on LRA but using reload those addresses are
    presented to it which chokes on them.  While reload is
    going away the change to make them work doesn't really hurt
    (and generally seems useful, as MULT and ASHIFT really are
    no different).  So just add it.

            PR target/116413
            * final.cc (walk_alter_subreg): Recurse on AHIFT.

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[cvs-commit at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

--- Comment #21 from GCC Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Andreas Schwab <schwab@gcc.gnu.org>:

https://gcc.gnu.org/g:a83e519ab2d4e7df2756411cd9d21c6f1b583429

commit r15-3233-ga83e519ab2d4e7df2756411cd9d21c6f1b583429
Author: Andreas Schwab <schwab@linux-m68k.org>
Date:   Tue Aug 27 21:01:00 2024 +0200

    m68k: Accept ASHIFT like MULT in address operand

    When LRA pulls an address operand out of a MEM it caninoicalizes a
    containing MULT into ASHIFT.  Adjust the address decomposer to recognize
    this form.

            PR target/116413
            * config/m68k/m68k.cc (m68k_decompose_index): Accept ASHIFT like
            MULT.
            (m68k_rtx_costs) [PLUS]: Likewise.
            (m68k_legitimize_address): Likewise.

[Bug target/116413] [LRA] [M68K] ICE: unrecognized insn in lra_set_insn_recog_data, at lra.cc:1036

[schwab@linux-m68k.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116413

Andreas Schwab <schwab@linux-m68k.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
         Resolution|---                         |FIXED
             Status|UNCONFIRMED                 |RESOLVED

--- Comment #22 from Andreas Schwab <schwab@linux-m68k.org> ---
Fixed for gcc 15.